Pattern recognition in lymphoid malignancies using CytoGPS and Mercator

Zachary B. Abrams, Dwayne G. Tally, Lin Zhang, Caitlin E. Coombes, Philip R. O. Payne, Lynne V. Abruzzo, Kevin R. CoombesBMC Bioinformatics 2021 · DOI

Introduction

In oncology, karyotyping remains the gold standard for diagnosing blood cancers. However, chromosomal records are traditionally stored in the International System for Human Cytogenomic Nomenclature (ISCN), a complex, semi-structured text format. This text-heavy nature makes large-scale computational analysis difficult, often requiring manual classification by pathologists. To address this, researchers developed a computational pipeline using two specialized tools: CytoGPS and Mercator.

The Computational Workflow

The proposed workflow transforms messy, text-based karyotypes into structured binary data suitable for unsupervised learning.

CytoGPS: The Translator

CytoGPS parses ISCN text strings and converts them into a binary "Loss, Gain, and Fusion" (LGF) model. This step effectively translates medical jargon into a vector of 1s and 0s, representing specific chromosomal events across the genome.

Mercator: The Analyzer

Mercator is an R package designed for unsupervised pattern recognition in binary data. A key feature is its use of the Jaccard distance metric. Unlike standard Euclidean distance, Jaccard ignores "negative matches" (0-0), which is essential for sparse genetic data where the absence of an abnormality in two patients does not imply biological similarity.

The tool employs Partitioning Around Medoids (PAM) for clustering and provides multi-view visualizations, including t-SNE and Multi-Dimensional Scaling (MDS). To manage large datasets, a density-based down-sampling technique ensures that rare clusters are preserved in visual representations.

Statistical Validation and Results

The study analyzed 22,741 karyotype samples from the Mitelman Database, covering 47 hematologic malignancies. The initial 2,748 binary features were refined to 814 unique informative features.

Scree plot and Bayesian sensitivity analysis used to determine the optimal number of clusters and feature groups.

As shown in the figure above, the researchers used a scree plot (a) and Bayesian sensitivity analysis (b) to determine the optimal dimensionality of the data. This process identified 134 distinct sample clusters and 72 feature groups. Approximately 70% of these clusters were characterized by a dominant cytogenetic abnormality (present in more than 60% of cases), demonstrating high fidelity.

Key Biological Insights

The pipeline successfully "recovered" known cancer hallmarks, such as the t(8;14) translocation associated with Burkitt lymphoma. More importantly, it revealed previously unrecognized patterns, such as the frequent co-occurrence of chromosome 14 loss (-14) and chromosome 21 gain (+21). This recurrent complex pattern had not been formally recognized as a separate entity by hematopathologists.

Visual analysis using t-SNE revealed "eye-shaped" clusters—ovals with a central point—representing groups of patients with nearly identical karyotypes. This provides a potential visual diagnostic tool for common genetic signatures.

Conclusion and Future Directions

While the pipeline identified 134 clusters, the true number of biological subtypes may lie between this figure and the 47 currently recognized morphologies. Future work will focus on clinical correlation, linking these genetic clusters to survival rates and treatment responses. The "data-type agnostic" nature of Mercator suggests it could be applied to various other "omics" fields beyond cytogenetics.

Summary updated

BibTeX
@article{abrams2021pattern,
  author = {Zachary B. Abrams and Dwayne G. Tally and Lin Zhang and Caitlin E. Coombes and Philip R. O. Payne and Lynne V. Abruzzo and Kevin R. Coombes},
  title = {{Pattern recognition in lymphoid malignancies using CytoGPS and Mercator}},
  journal = {BMC Bioinformatics},
  year = {2021},
  doi = {10.1186/s12859-021-03992-1},
}
Figure